Experimental Adaptation to Complexity

The pattern across five notes from four independent sources (Peter Palchinsky’s engineering principles via Tim Harford, Robert Yang and Tom Francis on game design, G.K. Chesterton’s fence heuristic, Dave Snowden’s Cynefin, and Donella Meadows-adjacent systems thinking) is a single claim about how intelligent agents should act when cause and effect are not predictable in advance: the right response to genuine complexity is designed, survivable experiments with tight feedback loops, preceded by understanding what kind of complexity you face, and qualified by not removing structures you don’t yet understand.

This idea is stable enough to name because it appears independently and from different directions in: the engineering framing (Palchinsky’s three principles: try new things, make it survivable, learn from feedback - (Experiment and design)); the learning systems design framing (wide partial-failure margins and tight feedback loops as the design conditions for productive learning - (Failure margins and feedback loops)); the epistemic prerequisite (understand before you change - Chesterton’s Fence as the complement to experimentation where reversibility is unavailable - (Chesterton fence)); the meta-diagnostic framing (Cynefin’s Complex domain as the territory where probe-sense-respond is the only valid approach, and the cliff edge as the danger of suppressing variation in the Simple domain - (Cynefin framework)); and the whole-system perspective (systems have shadow purposes; removing a subsystem without understanding its integrating role produces second-order failures - (Systems thinking)).

The internal structure of the idea has a natural hierarchy. First: diagnose whether you are in a domain where analysis can yield the right answer (Complicated) or where it cannot (Complex). The most costly error is treating Complex as Complicated - assuming expert analysis will yield the right plan when the system is adaptive and cannot be fully pre-understood. Second: in the Complex domain, vary (create diverse probes), contain risk (survivable scale, wide partial-failure margin), and learn (tight feedback, honest reporting). Third: before removing any structure (regardless of domain), understand what purpose it serves - the same impatience that motivates good reform also motivates hasty reform, and load-bearing functions are invisible precisely because the structure is doing its job.

The idea connects to Calibrated belief under uncertainty as the action complement: calibrated belief revision answers how to interpret evidence; experimental adaptation answers how to generate it and how to design action so that being wrong is survivable.